Reuters reconstructed Meta's plan to replace staff with AI, and the finding that killed it was that more code didn't become more product

More AI-generated code did not turn into proportionally more shipped product. That's the internal finding at the center of Reuters' August 26 special report, and it's the most useful number-free sentence I've read this month.

Project OT, for Organization Transformation, was hatched at Zuckerberg's January leadership retreat in Hawaii. The plan would have handed much of thousands of employees' daily work to AI, overseen by small "talent-dense" human cadres. It was scoped as two waves, May and November. Hours before the May 20 layoffs, Zuckerberg killed the second wave, and Meta stopped at a 10% reduction (Reuters).

Ars Technica reported the internal numbers behind the retreat. Project OT explored cutting or redeploying up to 60% of staff on some teams. Internal posts said unchecked AI agents were carrying out large-scale disruptive actions that "humans are unlikely to execute." Major technical and security incidents including service disruptions and possible data leaks rose 40% year over year, while time employees spent firefighting them rose 70% (Ars Technica).

Sit with that ratio. Incidents up 40%, firefighting time up 70%. The incidents got harder to clean up faster than they got more numerous. That's what it looks like when the failure mode changes character, not just frequency. A human who breaks production usually breaks it in a way another human can reason about. An agent that breaks production does something nobody would have thought to do, and the debugging starts from zero.

This is the strongest counterweight I've seen to this month's productivity claims, and it matters that it comes from inside a hyperscaler rather than from a vendor benchmark. Zalando reports 33% of PRs auto-approved. Warp's CEO says 30 to 35% of tasks automated weekly. Ramp says its own agent writes 75% of merged PRs. All of those are real. And Meta measured the same input surge and found the output didn't follow.

The reconciliation I'd offer, and I'm not confident in it: code generation was never the constraint on shipping product. Review capacity, integration risk, coordination, and the human judgment about what to build were the constraint. Remove the code-writing bottleneck and you pile work against the next bottleneck, which is people. The Faros AI telemetry Anthropic cited in its SDLC playbook fits this shape: across 10,000 developers and 1,255 teams, high-AI-adoption teams merge 98% more PRs, with review time up 91% and average PR size up 154%. Twice the merged PRs, twice the review time, and PRs two and a half times bigger. The work moved. It didn't disappear.

If you're planning headcount against an AI productivity assumption, Meta ran that experiment at a scale you can't and stopped. The measurement they used, shipped product rather than code volume, is the one to copy.